Modelling Left Ventricle Ejection Fraction Levels in Heart Failure Patients Using Circadian ECG Features and AI

Abstract

Heart Failure (HF) is increasingly recognized as a global pandemic, affecting more than 64.3 million people worldwide and causing significant disruptions in the normal functioning of their hearts. The estimation of left ventricular ejection fraction (LVEF) plays a crucial role in the diagnosis, risk stratification, treatment selection, and monitoring of HF (Preserved, Midrange, and Reduced). However, achieving a definitive assessment is challenging and typically requires the use of echocardiography. In contrast, electrocardiogram (ECG) is a relatively simple, quick, and cost-effective procedure that provides continuous monitoring of a patient’s cardiac rhythm compared to echocardiography. This thesis investigates the utilization of ECG for LVEF evaluation in HF patients through a thorough analysis of 24-hour ECG recordings and their derived features. Time- and frequency-domain characteristics (e.g., the P-QRS-T components of the ECG wave, the instantaneous frequency, and the spectral entropy) were extracted and used to train machine learning (ML) techniques, including deep learning (DL), for the estimation of LVEF levels and classification of the three categories of HF. LVEF levels were best estimated using rational quadratic Gaussian Process Regression (GPR) and fine decision tree regression models, with an average root mean square error (RMSE) of 3.83% and 3.42%, and correlation coefficients of 0.92 (p<0.01) and 0.91 (p<0.01), respectively. HF patients were best-classified using decision trees (TREE) and K-nearest neighbors (KNN) models, with an overall accuracy of 91.2% and 90.9%, and an average area under the curve of the receiver operating characteristics (AUROC) of 0.98 and 0.99, respectively.

Additionally, the analysis of the heart’s 24-hour circadian functionality (correlated with ECG characteristics) suggested that the time periods of midnight-1 am, 8-9 am, and 1011 pm contributed to the lowest RMSE values between the actual and predicted LVEF levels and the highest classification accuracy of the three HF categories. Further leveraging these findings and DL techniques, we selected the high-risk period from 10-11 pm as the ECG input for a hybrid model of Convolutional Neural Networks (CNN) and Long Short Term Memory (LSTM) layers to classify HF subtypes directly from raw ECG signals. This approach achieved an overall accuracy of 86%, with subclass accuracies of 89% for preserved, 80% for midrange, and 84% for reduced HF. However, ML models generally outperformed DL models due to the structured nature of ECG features and the relatively small dataset, which benefits from the interpret-ability and efficiency of traditional ML approaches over the higher complexity and data requirements of DL techniques. Finally, the proposed techniques for 24-hour ECG analysis in this thesis will help develop automated, low-cost, continuous, and effective ECG-based cardiac care systems that can achieve early diagnosis and personalized management of HF, ultimately improving patient outcomes and enabling real-time monitoring.

Date of Award	16 Dec 2024
Original language	American English
Supervisor	AHSAN Khandoker (Supervisor)